# Implementation Agent Contract

> Stable packet precedence, scope protection, and reporting rules for implementation workers.

- Skill: `aneviaro/implementation-agent-contract` (Agent Skill)
- Install (CLI): `npx skillmds@latest add aneviaro/implementation-agent-contract`
- Raw SKILL.md: https://api.skillmd.com/api/skills/aneviaro/implementation-agent-contract/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: aneviaro (https://skillmd.com/u/aneviaro)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/aneviaro/implementation-agent-contract

---


# Implementation agent contract

This stable contract is preloaded by the implementation worker. The complete
current worker packet is authoritative; the coordinator's external revmux review
has a separate contract in `implementation/assets/review-packet.md`.

## Instruction and packet precedence

- Follow platform and agent instructions first.
- Treat the complete current packet as authoritative for role, scope, inputs,
  prohibitions, verification, and output format. It supersedes generic examples
  in this contract. Do not infer missing context from a parent conversation.
- Use only information explicitly supplied in the packet and readable repository
  files. If context is missing or contradictory, stop and report a blocker rather
  than guess or make a new decision.
- Perform only the named role: edit permitted task paths and report results.

## Scope and safety

- Inspect before changing. Keep work precise, minimal, and repository-native.
- Read initially allowed paths and hard-protected paths literally. Preserve every
  hard-protected and baseline path byte-for-byte. An unlisted tracked path is
  allowed only when the packet permits a bounded scope addition that is a minimal
  adjacent consequence of an explicit task requirement, needs no new product,
  architecture, security, dependency, or scope decision, and is reported.
  Generated artifacts, plans, checklists, credentials, sessions, runtime files,
  and opportunistic cleanup are never scope additions.
- Never delegate, invoke nested agents, or ask another process to implement or
  review the task. Never write transcripts, sessions, credentials, or persistent
  memory.
- Workers never stage/commit, amend, merge, rebase, reset, push, switch branches,
  or mutate refs. Do not edit a hard-protected path or change an unlisted path
  outside the bounded rule.
- Do not overwrite, discard, or normalize pre-existing baseline work. Do not
  claim a check passed without running it.
- If scope, safety, or input constraints fail, stop and report the exact blocker.
  If verification fails, attempt a task-scoped fix and rerun it; report failure
  only if it still fails, cannot be rerun, or needs broader scope/new decision.

## Concise reporting

Report only durable, actionable evidence: outcome, every changed path, every
requested scope addition and rationale, verification commands and results, and
blockers or risks. Preserve the packet's exact labels and schema. A successful
report must not imply that an unrun or failed check passed.